Abstract:

Diffusion models are rapidly reshaping the paradigms of image and video generation, thus shifting generative artificial intelligence from a research-oriented tool toward a foundational infrastructure for digital-content creation. However, alongside the remarkable improvements in generative capability, a series of governance risks have emerged, including content safety, copyright ownership, fairness and bias, lack of truthfulness, and privacy leakage. Existing studies rarely focus on the intrinsic conflicts among different safety objectives: excessively strong safety alignments may degrade generation quality, constraints on truthfulness may suppress creative capability, and higher visual realism may increase the risks of deepfakes and privacy abuse. Hence, this paper systematically reviews the core governance risks of diffusion models from five perspectives: content safety, copyright and intellectual property, fairness and bias, hallucination and truthfulness, and privacy leakage. Furthermore, we propose an “alignment–quality–truthfulness” (AQT) trade-off framework to provide a unified perspective for analyzing the fundamental tensions and underlying mechanisms in diffusion models. We further argue that data memorization, concept entanglement, guidance amplification, the lack of world modeling, and data distribution bias jointly constitute the major sources of current governance challenges in diffusion models. The proposed AQT framework aims to provide a unified analytical perspective for diffusion-model safety research and facilitate the evolution of generative models from merely “high-quality generation” toward generation that is safe, truthful, and trustworthy.

Key words: diffusion models, trustworthy generation, model safety